Reliability and disease sensitivity are dissociable properties of EEG foundation-model representations

Abstract

Abstract EEG foundation models (EEG-FMs) are evaluated almost entirely on disease-discrimination accuracy. A clinical biomarker additionally requires measurement reliability, the stability of repeated measurements on the same individual, which regulatory biomarker frameworks treat as a prerequisite that discrimination does not imply. We asked whether frozen EEG-FM representations provide such stability, whether it is predictable from conventional model descriptors, and what information supports it. We measured test-retest reliability, disease discrimination, and representation distinctiveness for nine frozen representations: six EEG-oriented foundation models spanning masked, contrastive and predictive pretraining, handcrafted spectral features, and two general-purpose time-series models with no EEG exposure. All were evaluated under one preprocessing pipeline across two healthy retest cohorts, at roughly one month and two years, and three neurodegenerative cohorts. Reliability was measured in healthy adults only; the disease cohorts contribute cross-sectional discrimination. Reliability, quantified as the intraclass correlation coefficient (ICC), varied enormously (mean 0.08 to 0.76; coefficient of variation, CV, 53.0%) while disease discrimination, measured as area under the receiver operating characteristic curve (AUC), occupied a far narrower range across the same nine (AUC CV 5.4%), a roughly tenfold difference in relative dispersion, described rather than formally tested. We did not test formal AUC equivalence, so we describe discrimination as varying substantially less than reliability rather than as equivalent. The variation was not consistently explained by pretraining paradigm or domain among the models studied, and a model with no EEG exposure was among the most reliable tested. Alpha-band information contributed disproportionately to reliability, whereas theta-band information ranked first for Alzheimer disease and frontotemporal dementia discrimination, directionally consistent with established EEG evidence in both conditions. Only the reliability half of that contrast is individually significant. Subspace geometry and band ablation, two methodologically distinct analyses, both indicate that the two properties are partially, not fully, dissociable, and network architecture determines whether the dissociation is preserved, traded off, or jointly degraded across depth. Reliability showed no detectable association with discrimination, pretraining paradigm, or domain, and had to be measured directly. We recommend it become a standard evaluation axis for EEG-FM representations intended for longitudinal or biomarker use, and release a reproducible pipeline.